To achieve lifelong language learning, pseudo-rehearsal methods leverage samples generated from a language model to refresh the knowledge of previously learned tasks. Without proper controls, however, these methods could fail to retain the knowledge of complex tasks with longer texts since most of the generated samples are low in quality. To overcome the problem, we propose three specific contributions. First, we utilize double language models, each of which specializes in a specific part of the input, to produce high-quality pseudo samples. Second, we reduce the number of parameters used by applying adapter modules to enhance training efficiency. Third, we further improve the overall quality of pseudo samples using temporal ensembling and sample regeneration. The results show that our framework achieves significant improvement over baselines on multiple task sequences. Also, our pseudo sample analysis reveals helpful insights for designing even better pseudo-rehearsal methods in the future.

Lifelong learning (LL) aims to train a neural network on a stream of tasks while retaining knowledge from previous tasks. However, many prior attempts in NLP still suffer from the catastrophic forgetting issue, where the model completely forgets what it just learned in the previous tasks. In this paper, we introduce Rational LAMOL, a novel end-to-end LL framework for language models. In order to alleviate catastrophic forgetting, Rational LAMOL enhances LAMOL, a recent LL model, by applying critical freezing guided by human rationales. When the human rationales are not available, we propose exploiting unsupervised generated rationales as substitutions. In the experiment, we tested Rational LAMOL on permutations of three datasets from the ERASER benchmark. The results show that our proposed framework outperformed vanilla LAMOL on most permutations. Furthermore, unsupervised rationale generation was able to consistently improve the overall LL performance from the baseline without relying on human-annotated rationales.